<p><strong>TLDR:</strong> Meta launched Muse Code in beta today, a terminal coding agent built on Muse Spark 1.2 that scores 59.3% on DeepSWE 1.1 — competitive but behind Claude Code and GPT-5.6 Terra. The bigger story is Google DeepMind: Demis Hassabis is stepping back from day-to-day operations, and Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are all leaving to co-found Discovery Loop, a public benefit corporation backed by Radical Ventures, Khosla, and Alphabet itself. Prime Intellect also released Prime Agent, an open-source self-improving coding agent that scores 95.5% on ARC-AGI-3 using Claude Opus 5. Anthropic confirmed it is building custom silicon, and OpenAI published ten AI-generated mathematical breakthroughs using its internal Astra model.</p>
<hr>
<h2>Jeff Dean and three Gemini co-leads leave Google to found Discovery Loop</h2>
<p>Jeff Dean (27-year Google veteran, TPU architect, Gemini co-lead), Sanjay Ghemawat, Oriol Vinyals, and Quoc Le announced they are leaving to co-found Discovery Loop, a public benefit corporation aimed at automating the hypothesis-experiment-evaluation research cycle. Radical Ventures and Khosla Ventures are co-leading the seed round, with Alphabet participating as both investor and compute partner — giving Discovery Loop frontier-scale TPU access for at least a year. Demis Hassabis simultaneously stepped back from day-to-day DeepMind operations to become Chair and Chief Scientist of Alphabet, with CTO Koray Kavukcuoglu taking over product execution. Alphabet shares fell roughly 5% on the news. Four of the people who built the foundation of modern AI at Google all left on the same day — that’s not easy to frame as routine. <a href="https://daily.dev/posts/eysQpFbUp" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Meta ships Muse Code beta: 59.3% on DeepSWE, persistent agents, 24-hour GPU kernel run</h2>
<p>Meta launched Muse Code, a terminal coding agent built on Muse Spark 1.2, which was co-trained alongside the agent scaffolding rather than bolted on afterward. On DeepSWE 1.1 it scores 59.3%, beating Grok Build 4.5 and Gemini 3.6 Flash but sitting behind Claude Code with Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). The more interesting engineering detail: agents maintain persistent context across a full session and fan out to parallel sub-agents in isolated worktrees for large jobs, with every tool call hitting a local event log before executing so crashes resume from the last checkpoint. That durability is what let it run 1,000+ tool calls over 24 hours optimizing GPU kernels for NVIDIA Hopper. The CLI is in beta and reportedly doesn’t fully work yet. Pricing matches Spark 1.1 at $1.25/$4.25 per million input/output tokens. <a href="https://daily.dev/feed-by-ids?id=Mw1ZVVSLC&amp;id=5ny68Kwkc&amp;id=x0rc2iBej&amp;id=vpzz9s6iQ&amp;id=l30QjnNaW&amp;id=Mv6zRQro3" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Prime Agent scores 95.5% on ARC-AGI-3 using a self-modifying harness and recursive sub-agents</h2>
<p>Prime Intellect released Prime Agent, an open-source coding agent built around two ideas: the Recursive Language Model (RLM), which treats context as a variable and handles sub-agent delegation as function calls inside a persistent IPython REPL, and a Continual Harness that lets the agent create, update, and delete its own prompts, skills, and sub-agents at runtime. Using Claude Opus 5, it scores 95.5% on ARC-AGI-3 — just above the reported human expert baseline of 95.4%. Whether that qualifies under the benchmark’s official evaluation rules is an open question given the use of code execution, persistent memory, and recursive sub-agents, and some observers have flagged it. It also outperforms native harnesses on OOLONG, LongBenchPro, and EmulatorBench. Installs with a single curl command. <a href="https://daily.dev/feed-by-ids?id=sIAfIhERS&amp;id=O7WSsFAuY" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>OpenAI publishes ten AI-generated math breakthroughs using internal Astra model</h2>
<p>OpenAI published ten mathematical discoveries produced with its internal Astra model, accompanied by a 249-page research paper, a 62-page process explanation, and formal Lean 4 proofs on GitHub. Results span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics — including solutions to longstanding Erdős problems and newly constructed non-Sofic groups. The workflow was human-AI collaboration: Astra generated proofs, researchers refined them into papers, Astra translated results into Lean 4 for formal verification. Estimated inference cost was roughly $2,000 in API tokens. Each paper will undergo independent peer review, and the formal proofs let anyone verify logical correctness without trusting the lab’s word. <a href="https://daily.dev/feed-by-ids?id=V6XJXoFws&amp;id=WWmtmIGYt" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Anthropic confirms custom silicon program, hiring engineers who have ‘shipped silicon’ at $320K–$485K:</strong> Anthropic officially confirmed it is building an in-house chip team to design custom silicon for Claude, with job listings targeting engineers who have shipped silicon and salaries ranging from $320K to $485K — joining OpenAI (Jalapeño with Broadcom), Meta (Iris), and Mistral in the race for proprietary AI silicon, with custom chip development estimated to cost around $500 million. <a href="https://daily.dev/posts/arYTP8qlS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cursor open-sources Mixture-of-Kittens kernel, 41% throughput gain without new hardware:</strong> Cursor open-sourced MoK, a training kernel that fuses token dispatch, expert computation, and result gathering into one deterministic kernel that never pauses for CPU input, delivering a 41% throughput gain (760.9 to 1,070.2 tokens/sec across 512 GPUs) and 2.37x faster forward passes versus the strongest public baseline — all on existing NVL72 hardware. <a href="https://daily.dev/posts/91Jdz5PU2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Sapiom raises $35M Series A to route agent calls to cheapest capable model, Anthropic is a backer:</strong> Sapiom, which sits between AI agents and the models they call to enforce budgets before spend occurs, raised a $35M Series A led by Dragonfly (total funding $50M), with Anthropic as a notable backer — a customer case study shows Polsia cutting its Anthropic bill from $1.2M to roughly $100K per month after adoption. <a href="https://daily.dev/feed-by-ids?id=7ANr0wB4m&amp;id=l30QjnNaW&amp;id=7zxchxhGp" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>CISA adds Langflow CVE-2026-9198 (CVSS 9.8) to exploited vulnerabilities catalog — second Langflow flaw in a month:</strong> CISA added CVE-2026-9198, a critical unauthenticated remote code execution flaw in Langflow (IBM’s agentic AI framework) with a CVSS score of 9.8, to its Known Exploited Vulnerabilities catalog — the second Langflow vulnerability flagged in a month, giving federal agencies three days to apply mitigations. <a href="https://daily.dev/posts/ci76jfiSh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ninth Circuit rules Perplexity’s Comet AI shopping agent can legally access Amazon:</strong> The US Ninth Circuit overturned an injunction blocking Perplexity’s Comet agent from Amazon, ruling that because Comet acts on a user’s explicit instruction, it is legally the user accessing Amazon’s servers — killing Amazon’s CFAA claim and giving other AI agent developers a legal template to argue their tools aren’t unauthorized access. <a href="https://daily.dev/posts/tZuPNgBQd" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Meta launched Muse Code in beta today, a terminal coding agent built on Muse Spark 1.2 that scores 59.3% on DeepSWE 1.1 — competitive but behind Claude Code and GPT-5.6 Terra. The bigger story is Google DeepMind: Demis Hassabis is stepping back from day-to-day operations, and Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are all leaving to co-found Discovery Loop, a public benefit corporation backed by Radical Ventures, Khosla, and Alphabet itself. Prime Intellect also released Prime Agent, an open-source self-improving coding agent that scores 95.5% on ARC-AGI-3 using Claude Opus 5. Anthropic confirmed it is building custom silicon, and OpenAI published ten AI-generated mathematical breakthroughs using its internal Astra model.

Meta ships Muse Code, Jeff Dean leaves Google to found Discovery Loop

<p>four names walking out together is what reshapes the map here. that’s the MapReduce and AlphaFold lineage leaving one building. Alphabet backing the spinout reads as a hedge rather than a loss.</p>
